How managing NHIs can help teams secure AI agents

COMMENTARY: An AI agent spins up in a CI/CD pipeline job with broad access and no expiration date. It was created for a one-off experiment, but weeks later it’s still active, long forgotten. Eventually, it triggers a failure or its token leaks. When security investigates, there’s no audit trail, no clear owner, and no rollback plan.

So, what started as a productivity shortcut has become a potential security incident.

Consider the same scenario involving a human identity in a mature environment: all identities are discovered, and access gets scoped and time-bound. The system also monitors the behavior, and makes sure there's safe revocation and clear ownership. Governance isn’t bolted on after the fact—it’s designed in from the start.

[SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Read more Perspectives here.]

This contrast captures the reality many organizations face today. AI agents are woven into enterprise workflows, but without a way to manage them as governed identities, organizations risk turning efficiency gains into security liabilities.

Why non-human identities matter

Modern enterprise environments span containers, multi-cloud deployments, CI/CD pipelines, and increasingly, autonomous AI systems. In this world, non-human identities (NHIs)—service accounts, API tokens, and ephemeral AI agents—make up the vast majority of all identities.

Yet most IAM programs remain people-centric. Human logins are governed carefully, but machine and agent identities often sprawl unchecked. They are easy to create, hard to monitor, and rarely tied back to clear ownership. This results in identity debt at machine scale: agents with excessive permissions, long-lived tokens, secrets left unrotated, and blind spots where no one can say with certainty who—or what—has access.

A maturity model for AI agents

Today, security leaders manage all this sprawl by adopting maturity models that frame AI identity management as a staged journey. The path usually begins with visibility—knowing what agents exist and where they connect.

Next comes structured enablement, where onboarding and access controls are standardized. From there, organizations reach operational governance, applying policies, telemetry, and monitoring consistently at scale. Finally, teams reach autonomous action with control, where agents act in real time, but only within a trusted, auditable framework.

This progression reflects a simple truth: While it’s inevitable that we'll all adopt AI agents, this unmanaged adoption is unsustainable. CISOs must align AI rollout with a maturity curve that balances innovation with security.

Consider a supply chain agent that autonomously reroutes shipments. If its OAuth token gets stolen, attackers could manipulate logistics at scale. Or take a CI/CD pipeline agent that was granted broad administrative access: if its credentials leak, the entire software delivery process could be compromised.

These scenarios raise recurring questions:

Without answers, incident response slows, and accountability vanishes. In the world of AI agents, where activity happens at machine speed, that lack of visibility is a liability.

Identity as data

Teams need to treat AI agent identities as a data problem, not just a policy problem. Continuous visibility across NHIs, combined with ownership and behavior tracking, allows proactive governance. Effective AI agent identity management requires:

This data-driven approach shifts identity from static access control to dynamic assurance.

Today, many enterprises are still in the early stages—AI agents spin up ad hoc, with little oversight, often tied to experiments that drift into production. But adoption has accelerated: more organizations embed AI into core workflows that connect into business-critical systems, services, and applications, and regulators are beginning to demand accountability. Those who move early will win, adopting identity-first practices for agents before the sprawl becomes unmanageable while promoting speed and agility.

A secure future for AI depends on treating agents as the NHIs they are. That means discovery, ownership, behavioral monitoring, and automated lifecycle management—all reinforced by a maturity model that evolves with adoption.

Agentic AI promises enormous gains in productivity and automation, but it also introduces new risks. To secure agentic agents, we must make identity the control plane. By building visibility, ownership, and governance into the lifecycle of AI agents, organizations can innovate without opening the door to exploitation.

Itamar Apelblat, chief executive officer, Token Security

SC Media Perspectives columns are written by a trusted community of SC Media cybersecurity subject matter experts. Each contribution has a goal of bringing a unique voice to important cybersecurity topics. Content strives to be of the highest quality, objective and non-commercial.